Source-linked AI summary
ErgoAssist: Cognition-Aware Posture Feedback in Wearable Ergonomic Systems
Sarmistha Sarna Gomasta, Bhawana Chhaglani, VP Nguyen, Prashant Shenoy
TL;DR
ErgoAssist addresses disruptive posture feedback by combining IMU-based posture sensing with EEG-based cognitive-load estimation. Its cognition-aware alerting reduced alert frequency while improving posture correction and perceived usability in a preliminary real-time deployment.
Problem
Existing ergonomic systems often rely on posture thresholds, while prolonged device use is associated with poor posture, discomfort, and fatigue.
Method
ErgoAssist integrates IMU-based posture sensing with consumer-grade EEG estimation of task-induced cognitive load to schedule context-sensitive alerts.
Results
81% fewer alerts accompanied 38% higher posture correction and 43% better perceived usability under cognition-aware alerting.
Takeaways & Limitations
Modeling holistic user state enables fewer, better-timed ergonomic interventions without sacrificing ergonomic effectiveness.
Takeaways & Limitations
Long-term wearability, social acceptability, signal robustness, behavioral change, and adherence remain open challenges, with evaluation limited to short-term controlled tasks and a small preliminary deployment.
Abstract
from arXiv · showhide
Prolonged digital device use has made poor posture and musculoskeletal discomfort pervasive among knowl- edge workers. Existing ergonomic wearables rely solely on posture thresholds, frequently interrupting users during high-focus moments and leading to alert fatigue and abandonment. Yet posture and cognitive load are closely coupled, and most systems remain cognitively unaware. We present ErgoAssist, a head-worn ergonomic assistant that detects poor posture using IMU-based head tracking and estimates task-induced cognitive load using a consumer-grade EEG headband for continuous everyday use. In a controlled lab study, ErgoAssist achieves 81% posture classification and 90.2% task induced cognitive load estimation accuracy under leave-one-subject-out evaluation. In a preliminary real-time deployment, cognition-aware alerting reduces alert frequency by 81%, improves perceived usability by 43%, task performance by 25%, and improves posture correction rate by 38%, delivering fewer but better-timed interventions rather than merely suppressing alerts.
1 Introduction
ErgoAssist addresses poorly timed posture alerts by combining IMU-based posture sensing with EEG-based cognitive-load estimation to schedule cognition-aware feedback.
- Existing posture systems alert on threshold violations regardless of cognitive state, interrupting high-focus work and contributing to annoyance, alert fatigue, and abandonment.
- ErgoAssist combines IMU-based neck-posture monitoring with wearable EEG estimation of task-induced cognitive load to adapt alert timing.Alerts can be delayed or suppressed during high focus and prioritized during low-load moments or natural breakpoints.
- The study evaluates posture-only and cognition-aware feedback in a controlled user study with 24 participants performing Rest, Stroop, and Numerical Calculation tasks.
- 90.2% task-induced cognitive-load classification accuracy and 81.4% posture classification were achieved under leave-one-subject-out evaluation.
- 81% lower alert frequency, 38% higher posture-correction rates, and 43% better perceived usability were reported compared to posture-only feedback.
2 Background and Motivation
The paper motivates cognition-aware ergonomic assistance because screen use and poor posture are prevalent, while threshold-only alerts can arrive at disruptive moments.
- 2.2 Towards Cognition-Aware Ergonomic Assistance: The paper frames adaptive interventions as a way to promote postural habits without affecting performance.
- 2.1 Need for ergonomic assistant: A User Survey: 50 surveyed participants reported 7.36 hours/day of laptop use, 77.3% poor laptop postures, 93.2% eye strain, and 95.5% neck or upper-back pain.
- 2.1 Need for ergonomic assistant: A User Survey: Forward head posture increases cervical-spine load, with modest neck flexion accumulating into fatigue, long-term strain, and pain.
- 2.2 Towards Cognition-Aware Ergonomic Assistance: Posture thresholds evaluated independently of task demands or cognitive load can repeatedly notify users during high-focus moments and contribute to alert fatigue.
- 2.2 Towards Cognition-Aware Ergonomic Assistance: ErgoAssist fuses IMU posture sensing with EEG cognitive-load estimation to suppress or defer alerts during high focus and prioritize feedback during natural breakpoints.
3 Related Work
Prior work has advanced posture sensing, cognitive-state inference, multimodal sensing, and interruptibility-aware feedback largely as separate areas. ErgoAssist connects them to schedule ergonomic interventions.
- Few systems integrate wearable posture and cognitive sensing to study the interplay between mental workload and physical posture.
- Vision-based posture sensing offers high spatial fidelity but can be computationally intensive, privacy-sensitive, and vulnerable to occlusion or low light.
- Lightweight wearable EEG systems support inference of cognitive states including mental workload, fatigue, and attentional engagement in lab and field settings.
- Prior multimodal studies report benefits from combining IMU with EEG or EEG, ECG, and EMG for activity and fatigue recognition, especially under multitasking conditions.
- ErgoAssist integrates IMU posture sensing and EEG cognitive-load estimation to address when to intervene, targeting fewer but better-timed ergonomic corrections.
4 ErgoAssist Architecture
ErgoAssist combines head-worn IMU and EEG sensing with signal processing and machine learning to estimate posture and cognitive state for cognition-aware feedback. Its pipeline calibrates and cleans multimodal signals, extracts posture and EEG features, and fuses aligned representations for lightweight classification.
- System overview: ErgoAssist combines a head tracker and EEG headband to provide cognition-aware posture assistance during everyday computer use.The system tracks posture and cognitive state while exploring relationships between physical and physiological signals.
- Signal processing: IMU and EEG signals are calibrated, cleaned, processed independently, and fused at the model level for posture and cognition-related inference.
- System overview: The prototype captures EEG and IMU signals and runs pretrained sensor-based machine learning models in real time to infer cognition-aware alerts.
- IMU preprocessing: Personalized baseline normalization centers each user’s neutral posture near 0° and clarifies flexion ordering across D1, D2, and D3.The normalized distributions retain some overlap but support binary and three-class posture classification.
- EEG preprocessing: Synchronized IMU motion masks identify contaminated EEG segments before band-pass filtering, overlapping windowing, and rejection of heavily contaminated windows.Windows with more than 40% motion-contaminated samples are discarded.
- Feature extraction and fusion: Matched 5 s windows align EEG and IMU features, concatenate modality-specific vectors, and aggregate window-level features for downstream classifiers.
4.4 ML Inference
ErgoAssist uses lightweight traditional classifiers for posture, cognitive-load, and discomfort inference, selecting Random Forest for real-time alert policies. Nested validation balances predictive performance with computational constraints and addresses leakage and class imbalance.
- Model Validation: F1=.82 for grid-selected SVM versus F1=.78 for the selected setting in IMU binary classification, while larger gaps appear in multiclass settings.The grid search uses nested LOSO validation and evaluates classifier hyperparameters across specified regularization, kernel, tree, and boosting ranges.
- Perceived Discomfort Classification: Multimodal discomfort classification uses nested cross-validation, top-K feature selection, training-only standardization, and training-fold-only SMOTE.These choices target limited data, feature dimensionality, class imbalance, and prevention of test-set leakage.
- IMU-Based Posture Classification: Posture classification compares RBF SVM, Linear SVM, Random Forest, and Gradient Boosting on IMU features for multiclass and binary tasks.The models are chosen as a tradeoff among accuracy, interpretability, and computational efficiency for real-time use.
- EEG-Based Task-induced Cognitive Load Classification: EEG-based task-induced cognitive-load classification applies the same four model families to spectral and other EEG-derived features.Balanced weighting and shallow or regularized configurations are used to manage noisy, high-dimensional features and cross-subject overfitting.
- Model Selection: Random Forest is deployed for both posture and task-induced cognitive-load inference because it provides stable performance under 80/20 and LOSO evaluations.SVM, GB, and Linear SVM remain comparative baselines; model ensembles are avoided because their overhead conflicts with continuous on-device deployment.
5 Experimental Setup and Data Collection
The controlled IRB-approved study collects multimodal data from 24 participants performing rest and active cognitive tasks across three screen-distance posture conditions. Personalized posture calibration and real-time alert algorithms use IMU and EEG streams to support evaluation.
- Participants: Twenty-four participants complete approximately 20–25-minute sessions involving posture conditions and cognitive tasks in a controlled laboratory study.Participants are adults who report daily electronic-device use and provide informed consent.
- Posture Conditions: The protocol defines D1 as neutral posture at 70 cm, D2 as mild forward flexion at 50 cm, and D3 as deep forward flexion at 30 cm.Viewing distance is controlled instead of imposing fixed angular targets because natural head alignment varies across participants.
- Activities: Participants complete Rest and either Stroop or numerical-calculation activity to establish baseline and task-induced cognitive-load conditions.Rest lasts approximately one minute, while the active task window lasts three minutes; Stroop and calculation are established cognitive-load paradigms.
- Alert Scheduling: Posture-only alerts trigger after sustained non-neutral head posture, whereas cognition-aware scheduling combines posture and EEG-derived cognitive state to modulate alert timing.The cognition-aware policy delays or suppresses alerts during high mental engagement and releases them at low-focus breakpoints.
- Measurements: The setup collects IMU-based head orientation and EEG signals while measuring subjective workload and discomfort with NASA-TLX and CMDQ questionnaires.Both questionnaires use 7-point Likert scales after each posture condition.
6 ErgoAssist Performance Evaluation
ErgoAssist evaluates posture, cognitive-load, discomfort, feature groups, and their relationships under within-subject and LOSO settings. Results support multimodal sensing and cognition-aware intervention, while the pilot's small sample and short duration constrain its conclusions.
- Classification Performance: 81% posture classification and 90.2% task-induced cognitive-load classification are achieved under LOSO evaluation.Binary posture performance reaches .79–.81 across models, while binary cognitive-load recognition reaches .902 for Random Forest.
- Classification Performance: .57–.61 LOSO accuracy for multiclass posture classification shows that distinguishing D1, D2, and D3 is more affected by subject-specific variation.D2 is consistently the least reliably identified posture class.
- Feature Group Ablation Analysis: Time-domain IMU features are the primary posture signal source, but combining all feature groups yields the strongest binary cross-subject performance.Other feature groups alone reach only .52–.56 LOSO binary accuracy, whereas the full set reaches .81.
- Feature Group Ablation Analysis: Spectral ratio EEG features generalize more stably than absolute power or periodic alpha features, while the combined feature set reaches .90 binary LOSO accuracy.Spectral ratios achieve .92 LOSO accuracy in the binary ablation, compared with .76 for absolute power and .59 for periodic alpha features.
- Posture, Cognitive Load, and Discomfort: Deeper neck flexion is associated with higher perceived strain, and significant task-related reductions occur in periodic or relative alpha power.A Kruskal–Wallis test reports H=23.76, p<.001 for posture effects on TLX.
- Perceived Discomfort: Multimodal discomfort models outperform unimodal approaches, with Random Forest reaching .58 LOSO accuracy across participants.The reported takeaway links posture and task-induced cognitive load to user-perceived discomfort measured by NASA-TLX and CMDQ.
7 ErgoAssist in the Wild: A Preliminary Real-Time Comparative Study
The preliminary real-time comparison evaluates posture-only and cognition-aware alerting across 13 participants. Cognition-aware alerting substantially reduces interruptions while maintaining or improving response, usability, and task-performance outcomes.
- Alert exposure: 81% fewer alerts were issued under P+C than P-only, decreasing from 9.6 to 1.8 alerts across 13 participants.The reduction was consistent across known and unseen participant groups, while P+C alerts continued to exceed the pitch-deviation threshold.
- Posture response: 78.4% lower alert exposure accompanied an increase in mean participant-level response rate from 61.9% to 100.0% in five follow-up participants.Response rate is computed as corrected alerts divided by total issued alerts; P+C delivered equal or higher response rates per alert.
- Alert policy: P+C reduced alert exposure without sacrificing ergonomic outcomes, including for participants already achieving high P-only response rates.For example, P9’s alerts decreased from 16 to 1 with no loss in correction outcome.
- Alert policy: The 30 s override preserves intervention when prolonged poor posture occurs without a low-cognitive-load interruptible window.This creates a bounded tradeoff between reducing interruptions and maintaining ergonomic safety.
- Task performance: Mean arithmetic task scores increased by 8.2 points under P+C, from 33.2 (SD=18.9) to 41.3 (SD=21.8).Twelve of 13 participants achieved higher arithmetic scores under the cognition-aware condition.
8 Discussion
ErgoAssist frames posture assistance as an adaptive intervention problem by combining physical posture sensing with physiological cognitive-load sensing. The discussion supports reduced alert burden and improved user experience, while identifying substantial limits in wearability, signal fidelity, study duration, and personalization.
- Multimodal sensing: The cognition-aware policy combines IMU posture detection with EEG cognitive-load estimation to schedule alerts during less disruptive moments.The multimodal design treats IMU and EEG as complementary signals for physical strain and task-induced cognitive load.
- User experience: The real-time study reports 81% fewer alerts while improving perceived usability and maintaining ergonomic benefits.The discussion characterizes the integration as reducing unnecessary interruptions without sacrificing posture-related outcomes.
- Practicality and limitations: Consumer-grade EEG supports broad task-induced cognitive-load contrasts but is not suited to precise neurological assessment or fine-grained cognitive-state distinctions.Higher-density channels are typically required for fine-grained distinctions and source-level analyses.
- Practicality and limitations: The evaluation uses short-term controlled tasks, leaving long-term adherence, behavioral change, and naturalistic postural drift unresolved.The authors call for longitudinal studies of posture and cognitive load during extended real-world work periods.
- Future directions: The fixed cooldown and override thresholds cannot adapt to individual postural habits or cognitive-engagement patterns.Future work proposes learning personalized timeout and alert-deferral parameters from longitudinal observations.
9 Conclusion
The conclusion presents ErgoAssist as a head-worn system that combines IMU posture sensing with EEG-based cognitive-load estimation for cognition-aware ergonomic interventions. Its reported real-time results favor fewer, better-timed alerts while preserving ergonomic effectiveness.
- System contribution: ErgoAssist integrates IMU-based posture sensing with EEG-based task-induced cognitive-load estimation in a head-worn wearable system.The system uses this multimodal state representation to enable cognition-aware ergonomic interventions.
- Key findings: 81% fewer alerts and 38% better posture correction were reported for cognition-aware alert delivery than threshold-based approaches.The conclusion characterizes the result as fewer, better-timed interventions without sacrificing ergonomic effectiveness.
- Future work: Future longitudinal deployments will evaluate long-term adherence and real-world performance.The conclusion identifies these as directions for assessing the system beyond the reported study.
A Participant Demographics
The paper summarizes demographic characteristics for all 24 participants, including age, gender, device usage patterns, and vision range.
- Participant demographics: Demographic information was collected for all 24 participants.The summary includes age, gender, device usage patterns, and vision range.
- Participant demographics: Table 18 presents the participants’ demographic details.The table is identified as the source for participant demographics.
B Questionnaire Items
The study uses questionnaires to measure subjective workload, musculoskeletal discomfort, and perceived usability of the posture-feedback app. Items capture workload dimensions, discomfort across body regions, and reactions to alert use.
- Subjective Workload: NASA-TLX rates mental demand, physical demand, temporal demand, performance, effort, and frustration on a 1–7 scale.
- Musculoskeletal Discomfort: CMDQ rates discomfort in the neck, upper back, arms, and eyes from 1, indicating no discomfort, to 7, indicating extreme discomfort.
- Perceived Usability: The modified SUS assesses whether the app is desirable for frequent work use and unnecessarily complex.
- Perceived Usability: Usability items assess alert comprehensibility, technical-support needs, learning speed, confidence, and preparation required before use.
- Perceived Usability: The modified SUS also assesses inconsistency in alert timing and awkwardness of using alerts during tasks.
C Detailed Per-Alert Posture Correction Records
The detailed records report per-alert posture correction for unseen participants, including pitch changes, recovery time, and total corrections. Baselines are calibrated separately for P-only and P+C sessions.
- Table 19 covers per-alert posture correction for unseen participants P9–P13.
- Records compare baseline pitch with head pitch at alert onset and after correction.
- Recovery time measures seconds from alert to correction; 0.00 means immediate correction, while dashes indicate no correction within 30 seconds.
- Total corrected counts alerts resulting in correction against the total number of alerts issued.
- Baseline pitch values differ between P-only and P+C sessions because personalized calibration is recomputed independently before each session.